Code and main outputs for the controlled MRI audit illustration in “Stability Does Not Imply Evidence”

This repository contains the code, documentation, and main output files for the controlled MRI audit illustration used in the manuscript “Stability Does Not Imply Evidence: A Formal Analysis of Evidential Dependence in Attribution Based Medical XAI.” The release supports a paper facing reproducibility workflow for a controlled synthetic on real background experiment designed to illustrate the distinction between descriptive stability and audited decision dependence.

The code implements a unified observational construction on 2 D grayscale brain MRI backgrounds, in which two correlated synthetic factors are introduced. One is a declared lesion related target factor. The other is a spatially separate shortcut marker. The binary label is defined by the declared target factor. Two logistic regression predictors are included, namely a target dependent model and a shortcut dependent model. The release supports the main experimental components reported in the manuscript, including declared target region audit, model specific own candidate audit, stability only descriptive evaluation under mild non target perturbations, graded intervention analysis, and suppression based analysis.

This dataset release is intended as a controlled constructive illustration for methodological reproducibility. It is not a clinically realistic benchmark, does not provide clinical validation of true tumour evidence, and should not be interpreted as full empirical closure of the stronger audited evidential dependence framework discussed in the paper.

The release includes source code, detailed documentation, reproducibility instructions, and, where applicable, a version with main paper facing output files. Main materials and exploratory materials are explicitly separated so that the core manuscript evidence chain can be reproduced and inspected clearly.

The release package does not directly redistribute the original MRI background images. Background data preparation is described in the accompanying documentation, including instructions for preparing compliant grayscale MRI slice inputs before running the main workflow.

This repository is provided as the official submission time reproducibility release accompanying the manuscript.